TinyRTA

A real-time audio spectrum analyzer (RTA) built on the STM32G030F6P6 — an 8 KB RAM, 32 KB flash, Cortex-M0+ microcontroller with no FPU. TinyRTA samples ambient sound through a microphone front-end, runs a fixed-point FFT, and drives a 32x8 MAX7219 LED matrix with a 32-bar spectrum visualizer, all within the constraints of one of the smallest chips in the STM32 lineup.

Features

• Real-time audio sampling via timer-triggered ADC + circular DMA (ping-pong buffering, no interrupts)
• 8x hardware oversampling as a crude anti-alias boxcar filter
• 256-point fixed-point (q15) radix-2 FFT, no CMSIS-DSP; twiddle/window tables built at compile time
• Hann windowing for reduced spectral leakage
• Log-spaced grouping of FFT bins 2..127 into 32 bands, dB scale, peak-hold bars
• MAX7219 4-in-1 32x8 LED matrix output (bit-banged SPI)
• ~8 KB flash, ~2.2 KB RAM for the signal chain

Hardware

Component	Notes
MCU	STM32G030F6P6 (TSSOP20, 64 MHz Cortex-M0+)
Microphone	Electret mic + preamp (e.g. MAX9814 / MAX4466)
Display	MAX7219 4-in-1 32x8 LED matrix (FC-16 style)
Power	3.3V supply

No onboard mic preamp - an external analog front-end is required to bring the mic signal into the ADC's input range.

Wiring

Signal	STM32G030F6P6	Notes
MAX9814 OUT	PA0 (pin 7), ADC_IN0	~1.25 V DC bias, AGC regulates to ~1.4 Vpp. Optional RC anti-alias (1 kOhm + 47 nF)
MAX9814 VDD / GND	3.3 V / GND	Same rail as the MCU
MAX9814 GAIN	-	Unconnected = 60 dB, GND = 50 dB, VDD = 40 dB
MAX9814 AR	-	Unconnected = attack/release 1:4000
MAX7219 DIN	PA7 (pin 14)
MAX7219 CLK	PA5 (pin 12)
MAX7219 CS	PA4 (pin 11)
MAX7219 VCC / GND	5 V / GND	See App/max7219.cppm for 3.3 V logic caveats

Anti-alias filter (recommended)

The ADC samples at 8 kHz, so it can only represent content below 4 kHz. The MAX9814 output reaches up to ~20 kHz, and anything above 4 kHz folds back (aliases) into the displayed range as bars that are not really there. The on-chip 8x oversampling averages samples over each 125 us period, which only helps a little: it nulls 8 kHz and 16 kHz but lets 5 kHz through at about -6 dB.

Add a first-order RC low-pass between MAX9814 OUT and PA0:

MAX9814 OUT ---[ 1 kOhm ]---+---> PA0
                            |
                          47 nF
                            |
                           GND

The cutoff is fc = 1 / (2 pi R C) ~ 3.4 kHz. It rolls off gently (-6 dB/octave), so it attenuates the top bars a bit as well; a steeper second stage (another 1 kOhm + 47 nF) can be added if high-frequency content is strong.

Signal Chain

MAX9814 -> ADC (TIM3 64 kHz trigger, 8x oversampling = 8 kHz, circular DMA)
        -> DC removal -> Hann window -> q15 FFT (256-pt) -> |X|^2
        -> 32 log-spaced band sums -> dB -> peak-hold bars -> MAX7219

• Sample rate: 8 kHz (Nyquist = 4 kHz, sufficient for a visual spectrum display)
• FFT size: 256 points, 31.25 Hz bins; bands span 62.5 Hz .. 4 kHz (the lowest 18 bands are one bin each)
• Frame rate: ~31 fps (one spectrum per 256-sample block)
• Level scale: 6 dB per LED row, 48 dB total; see Bring-up and Tuning

Bring-up and Tuning

Work through these steps in order. Each one checks one more part of the chain, so a failure points at a single part.

1. Display

Set APP_DISPLAY_DEMO to 1 in Core/Src/main.c and flash. You should see an animated fake spectrum. If the image is upside down or mirrored, set kRotate180 in App/max7219.cppm. Set APP_DISPLAY_DEMO back to 0 afterwards.

2. FFT and bars, without the microphone

Set kTestTone = true in App/rta.cppm. This replaces the ADC data with a synthetic 1 kHz tone at a quarter of full scale. Bar 21 (counting from 0 at the left) should be lit almost to the top, with the rest dark or nearly dark. Set kTestTone back to false afterwards.

3. ADC input, in the debugger

Break in Rta::step() and inspect the analyzer object ((anonymous namespace)::rta_instance):

• adc_.buf_ should hold values around 1550 (the MAX9814's 1.25 V bias) in a quiet room and swing widely when you clap. Values stuck at 0 or 4095 mean a wiring or supply problem.
• adc_.overruns_ should stay at 0. If it grows, something in the main loop is blocking for longer than 32 ms.

4. Frequency check

Play pure tones from a phone tone-generator app close to the mic. The peak should appear at these bars:

Tone	Bar (0 = left)
100 Hz	1
440 Hz	12
1 kHz	21
3 kHz	29

The bars are 31.25 Hz apart below ~600 Hz (bars 0..17), then get wider on a log scale up to 4 kHz.

5. Levels

Tune the hardware gain first, then the software scale:

• MAX9814 GAIN pin: unconnected = 60 dB for room sound and speech, GND = 50 dB, VDD = 40 dB for loud music or a mic close to a speaker. The chip's automatic gain control then keeps the output near 1.4 Vpp for loud input.
• kTopQ3 (App/rta.cppm): the band power that fills a column to the top. One unit is ~0.38 dB, so +8 means ~3 dB louder input is needed to reach the top. Raise it if bars are pinned at the top during normal music; lower it if nothing reaches the upper rows.
• kDbPerRow (App/rta.cppm): dB per LED row, 6 by default (48 dB range). Use 3..4 for a livelier, jumpier display, 8 for a calmer one that shows quiet and loud parts together.

A good setting: in a quiet room only the bottom row flickers occasionally, and with music playing the bars move through the whole height, touching the top only on peaks. If room noise lights several rows, raise kTopQ3 or reduce the GAIN.

6. Look and feel

• Bar dynamics (App/bar_graph.cppm, shared with the demo): kDecayPerFrame sets how fast bars fall (larger = faster), and kPeakHoldFrames / kPeakFallPerFrame set how long the peak dot hangs and how fast it drops. Frames are ~32 ms.
• Brightness: the display_.init(4) argument in App/rta.cppm, 0..15.

Memory Budget

Buffer	Size	Location
ADC DMA ping-pong buffer (2 x 256 x u16)	1 KB	RAM
FFT work buffers (re + im, 256 x q15 each)	1 KB	RAM
Bar levels / peaks / hold (3 x 32)	96 B	RAM
Display framebuffer (32x8 mono)	32 B	RAM
Cosine table (twiddles, Hann window, test tone)	512 B	Flash

Total static RAM (incl. HAL, stack and heap reservations): ~3.7 KB of 8 KB.

Status

Signal chain implemented; level scaling still needs tuning on hardware (see Bring-up and Tuning).

Roadmap

• ADC + DMA acquisition pipeline
• q15 FFT
• Log-bucket grouping algorithm
• MAX7219 driver + bar rendering
• Amplitude log-compression / auto-gain for varying input levels
• Optional: adjustable sample rate / FFT size via config

License

TBD

Contributing

Issues and pull requests welcome once the initial pipeline lands.